aico 0.9.0

The learning release.

A correction you type becomes knowledge the agent keeps. Rate a reply ▼ and
say why; Remember this turns the note into a knowledge entry — a trigger
built from what you asked, your note as the guidance, filed with the project —
which the agent is shown on any later task whose wording matches. You read both
fields before they are kept. Nothing is adopted because the agent decided it
should be; that is the whole difference between a lesson and a confidently
wrong rule. In VS Code a ▼ asks What went wrong? immediately and shows the
trigger before you press Remember.

A skill can be measured, and improved only when it measures better.
aico skill eval <name> scores a skill against tasks with planted, checkable
answers; aico skill optimize <name> is SkillOpt's loop at personal scale — a
separate optimiser model reads the failing runs and proposes at most four edits a
step, a candidate is kept only if it scores strictly higher on validation tasks
the optimiser never saw, and rejected proposals are shown back so they are not
tried twice. Graders are regexes, files and tool-call counts, never a model
judging a model. It writes a draft for you to diff and register; it never
touches the shipped skill. Baseline on the cheap model, six tasks for eleven
cents: security-review 1.00, commit 1.00, init 1.00, review 0.92.

Fixed: auto-compaction had never run outside the terminal. A browser or
VS Code session could sit at 100% of a million-token window with the meter
saying so and nothing acting on it. The server now folds older turns at 75% of
the window, before and after a turn, and says so. Proven live: a 1,000-token
threshold, four turns, transcript pinned at three messages.

Fixed: both clients silently dropped the server's notices — "the agent you
addressed was deleted" was never shown. They render now.

Full notes: https://github.com/suhail-akhtar/aico/blob/main/CHANGELOG.md

The attached .vsix is the VS Code extension (0.6.1):
code --install-extension aico-vscode-0.6.1.vsix